How it was made, and what is wrong with it
This section was researched and written by a model, and checked by further runs of the same model. No person reviewed each line. This page says what was done, how much each check changed, what was left out and where the work is weak.
Research runs of one model read each vendor's own public pages (job listings, research posts, product pages and terms) and noted what each page shows the vendor being asked for or paid to do. A further run grouped the notes into signals and wrote each headline. Nothing was read from social media. This section was drafted by Claude, a model made by Anthropic. Anthropic buys from, sells to or competes with companies named here, so read what it says about them with that in mind. The rules every page of this section follows are in How to read this.
From candidates to dossiers
- 1425candidate sources of records, listed by separate passes that each looked through a different lens
- 803distinct kinds of record once the passes were merged, counting those that would have to be made; each was searched for anyone already selling or using it
- 41examined closely in a dossier
The kinds of record are on every kind of record; the dossiers are listed here.
What each check changed
Each check was a separate run that read the first run's work against its evidence. The share it changed is the best measure here of how far a first reading can be trusted.
| What was checked | Checked | Changed | What the check did |
|---|---|---|---|
| Vendors paired with signals | 1858 | 868 | A second run read every pairing of a vendor with a signal against the vendor's own note, and withdrew those no note supported. |
| Further vendors from a public register | 520 | 342 | Notes from the pages of further vendors, found through a public register of the trade, were read the same way; a reviewing run paired each with the signals it supports and left the rest unpaired. The register was used only to find each vendor's own pages. |
| Signals | 211 | 98 | The same runs reworded the signals whose headline, or stated implication, claimed more than the notes show. |
| Prices | 93 | 10 | A run read each price against the passage quoted for it, and rejected those where the passage stated a different figure or less than the amount claimed. |
| Levels of the kinds of record | 803 | 83 | Two runs read the level of every kind of record against the finding behind it, and moved those the finding did not support, most of them downward. |
| Dossier sources that would not open | 102 | 96 | Every dossier source that would not open was given another readable page or an Internet Archive copy, or the sentences resting on it were cut. |
| Numbers in the dossiers | 335 | 47 | Every number in the dossiers was tested against the page its sentence cites, and those that failed were rewritten to the page's own form, turned into words or removed. |
| Quotations in the dossiers | 69 | 14 | Every quoted passage in the dossiers was tested word for word against its cited page, and those that failed were corrected to the page's words or removed. |
What is left out
- Anything read only on social media. Several vendors announce their revenue there and nowhere else; those figures are not on this site.
- Prices that the page does not state in a single passage as this site's fetch sees it. Many more were found than are shown.
- Vendors none of whose pages answered a polite fetch, or whose pages supported no signal.
- What paid newsletters report, unless a page the site can read carries the same fact.
- How each kind of record might be obtained, what a business built on it might earn, and who to ask. Those notes are the owner's and are not published.
- Much of materials science. A first pass is in: the sellers it found are long-established crystal-structure databases, and laboratories that make new data keep it. Coverage of China, Japan and of battery and polymer test data is thin.
What is wrong with it
Three further runs were asked to attack the work. These are their findings that still stand.
- The first screen favoured the obscure. It marked a kind of record down for being already sold, so the markets where money plainly moves scored lowest. A second set of dossiers was written on those markets; the first ranking is not published.
- The passes that listed candidates were not independent. All were told the valuable shape was a decision followed by its result, so their agreement shows what is obvious and not what is good.
- Most levels rest on one search. In several dossiers a closer look found an incumbent the first search had missed, so a mark of nobody found is weaker than it looks.
- An early summary misstated several facts, among them a bid in a bankruptcy described as a completed sale. Each was corrected against the filing.
- No kind of record came out strong. No buyer was shown for a found record valued because its outcome is attached, which was the idea the work began with.